Tropical Cyclone Wind Speed Retrieval from SAR Imagery Using A Neural Network

Abstract

Tropical Cyclone (TC) wind speed retrieval from spaceborne synthetic aperture radar (SAR) cross-polarization acquisitions is a hotspot in SAR remote sensing application. In this paper, we collected 12 Interferometric Wide swath (IW) mode VH-polarization images observed by the European Space Agency Sentinel-1 satellites and the collocated Stepped Frequency Microwave Radiometer and Soil Moisture Active Passive radiometer wind speed observations up to 74 m/s under TC conditions. Matching samples were divided into two datasets. Dataset 1 contained 3851 samples and was used to develop a new geophysical model function (GMF). Dataset 2 contained 1282 samples and was used for validation. Each dataset included the parameters of VH-polarized normalized radar cross section (VH NRCS), wind speed and radar incident angle. A two-layer feed-forward network was utilized to generate a GMF named as the Sentinel-1 IW mode Wind Retrieval Model Based on Neural Network (S1IW.NN). VH NRCS and incident angle were inputs. Wind speed was output. The network outputs' related coefficient with respect to targets was 0.94 for Dataset 1. In order to validate the S1IW.NN model, we compared the S1IW.NN model's retrievals and those from a traditional empirical GMF S1IW.NR, based on the wind speed observations in Dataset 2. For the S1IW.NN model, the overall bias, correlation coefficient (Cor) and root mean squared error (RMSE) were −0.03 m/s, 0.93 and 3.72 m/s, respectively. For the S1IW.NR model, the overall bias, Cor and RMSE were −0.17 m/s, 0.92 and 4.61 m/s, respectively. In addition, we carried out a case study of the Hurricane Dorian for evaluating the proposed GMF visually and statistically. Results showed that the proposed GMF has a higher accuracy than the traditional model. Last but not least, the S1IW.NN model was able to retrieve the wind speeds at the sub-swaths' inner boundaries more smoothly, indicating the neural network's good performance in application to GMF regression.